How I Built Agent (2): ChatGPT Instead of a Textbook: My First Files and Mistakes
How I Built Agent — part 2 of 10.
When I started developing my own Agent, I had no development team and no finished architecture. I had an idea, an open ChatGPT window and a desire to find out what I could really create with generative AI. Those early weeks felt more like a practical course than a professional development project.
The chat explained; I uploaded the files
I described a problem and ChatGPT suggested an approach and produced some code. I saved the files, uploaded them to the server and tested the result. Then I returned to the conversation with another question. Sometimes one small adjustment was enough. Sometimes convincing-looking code did not work in my environment at all.
Those difficulties made me understand what a web server actually does, how components communicate and why configuration matters. I could simply have subscribed to a similar service. It would have saved time, but it would also have deprived me of the experience of discovering what goes on beneath a finished application’s interface.
A patient guide can still be wrong
One thing I enjoyed was being able to ask the same basic question over and over without embarrassment. But conversational learning has a weakness. A model may recommend an outdated method, miss an important condition or present an unverified assumption with complete confidence.
I learned to distinguish what AI tells me, what reliable documentation says and what I have actually verified. Those three things are not interchangeable. Confusing them is particularly risky when building something that needs to be dependable.
Learning in the opposite direction
I began with a concrete problem, followed it through conversation and experimentation, and only then sought the theory. This reverses the order of many textbooks. I would not claim that it is universally better. It did, however, lead me to questions that I might never have asked while simply reading.
The most useful outcome of those early weeks was therefore not a file on a server. It was a new way of learning. Generative AI can be an excellent guide as long as I continue to question and test its answers.
Next: How I moved from manual file uploads to Codex and GitHub.
